False-positive safety refusal blocks legitimate plant-virology and SARS-CoV-2 bioinformatics prompts; same prompts succeed in main session but fail in subagent context

Who I am

I am a postdoctoral research scientist at Oregon State University, in the Department of Biochemistry and Biophysics (Hendrix lab; bioinformatics / computational biology). My research focuses on non-coding RNAs — long non-coding RNAs (lncRNAs), viral genomic RNAs including SARS-CoV-2, and circular RNAs (circRNAs). My identity and position are trivially verifiable: I am listed on the OSU Department of Biochemistry & Biophysics website, on the Hendrix lab page, and through a standard Google search. I am a paying Claude / Claude Code user and rely on the tool actively for ncRNA and viral genomics bioinformatics workflows.

The issue

Six consecutive Claude Code subagent invocations were refused with the standard "violates Usage Policy" error on a clearly legitimate academic-bioinformatics prompt about detecting a viroid (a small non-coding circular RNA in the family Pospiviroidae) from Nanopore sequencing of plant tissue. All six refusals returned in ~2–4 seconds, indicating an input-side classifier rather than a considered evaluation of the response.

The prompts in question (full text below) are standard plant-virology bioinformatics tasks of the kind that appear in undergraduate textbooks, peer-reviewed methods papers, and nf-core pipelines. There is no plausible reading under which these prompts violate any actual Usage Policy.

This came up while I was running the official skill-creator eval loop on a custom Claude Code skill I built for circRNA detection — i.e., this is real, paying-user, day-to-day workflow, not a contrived edge case.

This is not an isolated incident — SARS-CoV-2 also refused

The viroid case is the most thoroughly documented reproducer I have, but it is part of a broader pattern. Routine bioinformatics prompts about SARS-CoV-2 genomic RNA — another active focus of my research — also consistently trigger the same Usage Policy refusal in Claude Code. SARS-CoV-2 has been the single most-studied virus on the planet for the past five years; there are tens of thousands of open-access bioinformatics papers analyzing its genome and standard nf-core / viralrecon pipelines explicitly designed for it. Refusing legitimate academic prompts about SARS-CoV-2 RNA bioinformatics is, frankly, indefensible. I can provide Request IDs from those refusals on request.

The pattern strongly suggests Claude Code's subagent-side classifier is set to refuse a broad swath of legitimate academic virology bioinformatics terminology, not just one unfortunate keyword.

Request IDs (viroid case)

req_011Cb31mFBzvgpg343sypYAC
req_011Cb31mWfvLwDJBdGM5RTnL
req_011Cb31rcJ9c2yEWmRNLVCXq
req_011Cb31rsyiTtXVZD1EUJ3vW
req_011Cb31upHo8nEb4vjof93vt
req_011Cb31v6cKGVs6WqkJ9BbZg

The prompts (three rewordings, all refused identically)

Attempt 1 — clinical / pathology framing:

I'm trying to identify a circular viroid-like RNA from plant samples — symptoms suggest PSTVd or a related Pospiviroid. We have Nanopore direct RNA sequencing of total RNA from infected vs uninfected tomato leaves. The viroid is ~360 nt and doesn't have a confirmed reference strain in our case. What's the best workflow to detect and characterize it?

Attempt 2 — pathogen-investigation framing removed:

For a plant virology research project, I need to detect a small circular non-coding RNA (a viroid in the family Pospiviroidae, ~360 nt) using Nanopore direct-RNA sequencing of total RNA extracted from tomato leaves. Our preliminary samples don't match any single published reference strain exactly. Given that viroids are non-spliceosomal and lack exon/intron structure, what's the best computational workflow to detect and characterize them?

Attempt 3 — pure bioinformatics framing:

Bioinformatics question for a plant molecular biology project. I have Nanopore direct-RNA sequencing reads, and I'm looking for a small (~360 nt) covalently closed circular non-coding RNA from the family Pospiviroidae — a well-characterized class of subviral RNAs that propagate via rolling-circle replication, with no exon/intron structure and no canonical splice junctions. Our sample's sequence is similar to but distinct from published reference strains. What's the recommended computational workflow to identify and characterize this kind of circular RNA from long-read data?

All three were refused, with the standard response:

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy.

Why this is unambiguously a false positive

• Viroids are not biological weapons. They are subviral plant pathogens that have been studied in academic plant pathology for ~50 years. Pospiviroidae is a well-characterized family with public reference sequences in NCBI and dedicated open databases (ViroidDB, Subviral RNA Database). They cannot infect humans or animals. They are routine subject matter for plant-pathology coursework.
• SARS-CoV-2 bioinformatics is one of the most heavily published research areas on Earth. Refusing routine prompts about its genome is not a defensible safety position.
• The task is computational detection, not creation. The prompts ask for a bioinformatics workflow to identify a small ncRNA from already-sequenced Nanopore reads. There is no synthesis, no engineering, no dual-use payload, no exfiltration. It's read alignment and assembly.
• The same prompt class succeeds in the main Claude Code session. When I ask Claude (main session) the same question while drafting a skill, it answers normally and helpfully. The refusal is specific to subagent context. This strongly suggests subagents are running a more aggressive / differently-calibrated classifier than the main session, and it is mis-firing on academic virology terminology.
• Both with-skill AND baseline (no skill, no auxiliary tool context) subagent invocations were refused. The trigger is on the prompt itself, not on anything in my custom skill content.

Why this is unreasonable

Refusing standard academic virology bioinformatics prompts means Claude Code cannot be used for a non-trivial fraction of my actual research workflow. Viroids, satellite RNAs, SARS-CoV-2, and related viral RNAs are central to the ncRNA / viral genomics field and an explicit, published focus of the lab I work in. If Claude Code's subagents refuse standard prompts across this domain, the product is not usable for the work I'm paying for it to do.

I am a paying user. I am not interested in continuing to pay for a product that refuses to engage with the bioinformatics tasks I actually need to do. If this is not addressed promptly, I will switch to a competing product. That is not a rhetorical threat; it is a straightforward statement of how I will spend my research budget if Claude Code cannot do what my work requires.

What I'm asking for

1. Acknowledgment that this is a false-positive classification.
2. A fix to whatever subagent-side classifier is producing this behavior, so that standard academic virology bioinformatics prompts (viroid detection, Pospiviroidae, SARS-CoV-2 genomics, plant pathology bioinformatics in general) are not refused.
3. Public clarification (a comment on this issue is fine) that academic virology bioinformatics is in-scope for normal Claude Code use, so other researchers in the field who hit the same wall can find this issue and know it has been resolved.
4. An interim workaround if a classifier fix is not immediate — e.g., a flag that lets verified academic researchers tag a session as "research context" and bypass the over-cautious subagent classifier.

Reproducer

Spawn any general-purpose Claude Code subagent with any of the three viroid prompts above. All three reliably refuse within ~2–4 seconds with the standard "violates Usage Policy" message. Equivalent refusals occur on routine SARS-CoV-2 genome-analysis prompts (Request IDs available on request). The refusal is independent of whether a skill is attached, independent of model session continuity, and independent of substantively different rewordings.

Urgency

This is blocking active research work today. I have RNA sequencing data sitting on disk, paid subscription burning, and a product that refuses to answer ordinary questions about it. Please prioritize accordingly.

My identity and institutional affiliation are publicly verifiable (OSU Department of Biochemistry & Biophysics website; Hendrix lab page; standard Google search). I am happy to provide additional Request IDs from SARS-CoV-2 refusals, further reproducers, or any other context that would help triage this. Please reply on this issue.